# Pilot retained — not the neutral-label experiment

Interrupted on 8 October 2026 after detecting that several condition labels revealed part of the withheld fact, for example “PHP not documented” and “not an automatic quote”. Even though the prompt said not to infer facts from labels, this is a design confound. The pilot’s outputs are not used to claim neutral missing-condition recognition.

45 responses were completed. One request was in flight when the process was stopped; its result and final bill are unknown. Known pilot cost: USD 1.86033 before tax. Its outstanding conservative reservation is USD 0.125813. This reservation remains charged against the cumulative USD 10 cap. Historical known spending before the pilot: USD 1.43470. Budget charged before the revised experiment: USD 3.420843 (known cost plus an upper bound, not an exact invoice).

The original protocol, corpus, requests, outputs and ledger remain intact. The revised experiment lives in ../attached/, uses neutral category labels, and keeps the C wording and retriever unchanged. This correction was made before any hosting-case pilot response was collected. It is a protocol correction, not exclusion of unfavourable results.